Multi-view Representations for Pose Invariant Face Recognition in Man and Machine

点击翻译按钮获取中文摘要

Humans have an incredible ability to identify familiar individuals despite substantial changes in the visual picture. These can include changes in lighting, viewing angle, distance, occluders and peripheral elements such as glasses and hair, and both static and dynamic shape change of the face due to speech and expression. Despite this highly functional human capability, facial appearance variations pose a significant challenge to computer vision systems. Achieving Pose-Invariant Face Recognition (PIFR) in computer vision systems remains a significant stumbling block to realizing the full potential of face recognition as a passive biometric technology. Extensive efforts have been made to try to solve the problem of pose-invariant face recognition in the development Artificial Intelligence, yet it remains a significant barrier. PIFR is achieved effortlessly by the human visual system but at present we do not understand the human system well enough to provide and implement plausible solutions to the clear technological challenges. The aim of this PhD studentship will therefore be to enhance our understanding of how human observers achieve pose invariant recognition of faces in order to inform AI strategies. We will particularly focus on multi-view or pose-aware strategies and compare these against object-based models or pose-agnostic approaches. The work will involve both extensive psychophysical experimentation with human participants and computational experiments exploring computer models of pose-invariant dynamic face recognition. The psychological and psychophysical experimentation will include behavioural studies as well as studies incorporating neuroimaging methods such as functional magnetic resonance imaging (fMRI) to help understand the brain networks involved in pose invariance in humans. These experiments will include determining whether, and how, human observers utilise or disregard the variations in the visual input whilst achieving pose invariant face recognition.